CN111949713A - Time series trend transformation inflection point prediction calculation method and device - Google Patents

Time series trend transformation inflection point prediction calculation method and device Download PDF

Info

Publication number
CN111949713A
CN111949713A CN202010919043.5A CN202010919043A CN111949713A CN 111949713 A CN111949713 A CN 111949713A CN 202010919043 A CN202010919043 A CN 202010919043A CN 111949713 A CN111949713 A CN 111949713A
Authority
CN
China
Prior art keywords
inflection point
short
term
slope
conversion
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202010919043.5A
Other languages
Chinese (zh)
Inventor
洪志令
吴梅红
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Individual filed Critical Individual
Priority to CN202010919043.5A priority Critical patent/CN111949713A/en
Publication of CN111949713A publication Critical patent/CN111949713A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • General Physics & Mathematics (AREA)
  • Development Economics (AREA)
  • Mathematical Physics (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Game Theory and Decision Science (AREA)
  • Fuzzy Systems (AREA)
  • Operations Research (AREA)
  • Probability & Statistics with Applications (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method for calculating a trend conversion inflection point of a time sequence. Firstly, designing a frame for calculating a time series conversion inflection point; on the basis, according to the historical trend of the time series, the experience is summarized, and a corresponding calculation strategy is designed and added into the framework. The method divides the output result into two typical operation types according to the range of the calculated data: a high frequency mode and a low frequency mode. After the two modes are trained respectively to obtain the optimal parameters, after data of each time sequence are updated every day, whether a conversion inflection point occurs on the time sequence on the day is calculated, and meanwhile, the condition that the conversion inflection point occurs on the tomorrow day is rapidly estimated by a binary search method. The method provides a new method for short-term prediction of the time sequence, improves the defects of the traditional prediction method and improves the accuracy.

Description

Time series trend transformation inflection point prediction calculation method and device
Technical Field
The invention relates to the technical field of data mining of time series, in particular to a calculation method for trend conversion inflection points of the time series.
Background
The time series is a set of data arranged according to a certain time interval, and the time interval can be any time unit, such as hour, day, week, month, and the like. If you are a shareholder, the stock price of a certain stock is a type of time series data that records the stock price of the stock at each point in time. If you are an operation and maintenance person, the monitoring data is a kind of time sequence data, for example, the monitoring data for the CPU of the machine, namely, the actual consumption value of the CPU on the machine is recorded at each time point. In addition, the method also comprises a power consumption time series, an electric merchant sales time series, an electroencephalogram time series, an air temperature time series and the like.
The analysis and prediction problem of time series is always a research hotspot. When time series analysis is used for prediction, a series of models are needed, and the models are collectively called time series models. When using time series models, it is always assumed that a certain pattern of data changes or a certain pattern of combinations occurs repeatedly. It is therefore necessary to first identify such patterns and then predict them by extrapolation.
According to different modeling theories, the prediction models can be divided into two major categories, namely a traditional fluctuation rate prediction model based on a statistical principle, which is popular and representative at present and comprises an ARCH model and an SV model, and an innovative prediction model based on a neural network, a grey theory, a support vector machine and the like. The two models have characteristics in prediction, but the prediction accuracy is still to be improved.
Disclosure of Invention
The invention discloses a method for calculating a trend conversion inflection point of a time sequence. Firstly, designing a frame for calculating a time series conversion inflection point; on the basis, according to the historical trend of the time series, the experience is summarized, and a corresponding calculation strategy is designed and added into the framework. The method divides the output result into two typical operation types according to the range of the calculated data: a high frequency mode and a low frequency mode. After the two modes are trained respectively to obtain the optimal parameters, after data of each time sequence are updated every day, whether a conversion inflection point occurs on the time sequence on the day is calculated, and meanwhile, the condition that the conversion inflection point occurs on the tomorrow day is rapidly estimated by a binary search method. The method provides a new method for short-term prediction of the time sequence, improves the defects of the traditional prediction method and improves the accuracy.
The method of the invention carries out calculation analysis after daily data updating, has no future function, and the generation of the subsequent conversion inflection point can not influence the result generated in the front. After the result of the conversion inflection point is visualized, the user can use the method intuitively and conveniently. Finally, the method is realized through a computer system, and a device is formed for predicting the time sequence input by the user.
The method comprises the following specific steps:
(1) designing a conversion inflection point calculation framework based on the long and short term lines;
(2) adjusting parameters for generating a long-term line based on a slope combination mode;
(3) according to the range of the calculated data of the short-term line, dividing the output result into two typical modes of high frequency and low frequency;
(4) generating a real prediction record according to the conversion inflection point by combining historical data, and calculating increase to perform parameter tuning;
(5) after the data of the time sequence are updated every day, calculating whether a conversion inflection point appears on the current day of the time sequence by combining the data of the current day; meanwhile, the condition of the inflection point of the switching on the tomorrow is estimated by a binary search method.
Wherein, in the step (1), a conversion inflection point calculation frame is designed based on the long and short term lines, and the method specifically comprises the following steps: two trend curves are calculated based on current and historical data: long term and short term lines. When the short-term line passes through the long-term line from bottom to top, recording the intersection point as an upper inflection point; conversely, when the short-term line passes through the long-term line from top to bottom, the intersection point is recorded as a lower inflection point. The long and short term lines are defined in a uniform manner as: line = EMA (SLOPE (D, K) × T + D, L). According to different values of the parameter variable K, T, L, the long and short lines are further specifically defined based on a unified form. As typically defined:
short-term line: ShortLine = EMA (SLOPE (D, K)b)*Tb+D,Lb) And = EMA (SLOPE (D,8) × 4+ D,14), i.e., after moving 4 steps forward with a fitting SLOPE of approximately 8 days as the trend inertia, performing an exponentially weighted moving average with a parameter of 14. In the formula Kb=8, Tb =4, Lb =14；
Long-term line: LongLine = EMA (SLOPE (D, K)s)*Ts+D,Ls) = EMA (SLOPE (D,20) × 10+ D,35), i.e. after moving 10 steps forward with a fitted SLOPE of approximately 20 days as the trend inertia, an exponentially weighted moving average with a parameter of 35 is performed. In the formula Ks =20,Ts=10,Ls=35。
Wherein, the step (2) adjusts parameters of the generation of the long-term line based on a slope combination mode, and specifically comprises the following steps: SLOPE (D, K) that primarily tunes long-term line LongLines)*TsAnd (4) partial. The adjustment to the long-term line based on the slope combination is as follows: the main objective is to adjust TsI.e. the number of steps taken forward. Note recent SLOPE SLP1= SLOPE (D,20), short-term SLOPE SLP2= SLOPE (D,4), ABS represents the absolute value, and there are four combinations depending on whether SLP1 and SLP2 are greater than 0:
(1) SLP1>0 and SLP2>0. At this time, if SLP2>= SLP1, then reduce TsA value of (d);
(2) SLP1>0 and SLP2<0. At this time if ABS (SLP2)>= SLP1, increase TsA value of (d);
(3) SLP1<0 and SLP2>0. At this time, if SLP2>= ABS (SLP1), increase TsA value of (d);
(4) SLP1<0 and SLP2<0. At this time if ABS (SLP2)>ABS (SLP1), then T is decreasedsThe value of (c).
In the step (3), the output result is divided into two typical modes of high frequency and low frequency according to the range of the calculation data of the short term line, specifically: the short term line is: ShortLine = EMA (SLOPE (D, K)b)*Tb+D,Lb) Further, two sets of related parameters are defined according to the range of the calculated data of the short term line: ultra-short term and short term. Ultra-short term line: ShortLineA = EMA (SLOPE (D,4) × 2+ D,7), when K is presentb =4, Tb =2, LbAnd =7, namely, the data move forwards by 2 steps according to the slope inertia of the data of nearly 4 days. Short-term line: ShortLineB = EMA (SLOPE (D,8) × 4+ D,14), when K is presentb =8, Tb =4, Lb=14, i.e. 4 steps forward by the slope inertia of the data of nearly 8 days. Combining the ultra-short term line and the short term line with the long term line, calculating an intersection point to obtain a conversion inflection point, and combining to form two modes: a high frequency mode and a low frequency mode. High-frequency mode: short-term line = ShortLineA; long-term line = LongLine; low-frequency mode: short-term line = ShortLineB; long-term line = LongLine.
Wherein, the historical data is combined in the step (4)Generating a real prediction record according to the conversion inflection point, and calculating and increasing to carry out parameter tuning, wherein the method specifically comprises the following steps: short-term wire shortlink = EMA (SLOPE (D, K)b)*Tb+D,Lb) And long-term line LongLine = EMA (SLOPE (D, K)s)*Ts+D,Ls) Are defined with respect to parameters including Kb，Tb，Lb，Ks，Ts，LsWhen the method is actually applied, parameter optimization is carried out by combining historical data of each time sequence; the method comprises the following specific steps: 1. taking a certain time sequence as an example, taking data of a period of time; 2. only one parameter is adjusted, and other parameters are fixed to generate a conversion inflection point; 3. generating a real prediction record according to the conversion inflection point, and recording the increase of the section from the occurrence of the upper inflection point to the occurrence of the lower inflection point; 4. calculating the times and the total growth condition of the inflection points of the conversion; 5. comparing with the last growth result, and reserving the result and the parameter with large growth; 6. repeatedly executing the steps 2-5 to obtain the optimal value of the current adjustment parameter; 7. the value of the adjustment parameter is fixed and the value of the other parameter is changed in a similar procedure. After all parameters have been adjusted, this is counted as one round. And (4) carrying out several rounds of parameter adjustment according to the steps, wherein finally obtained parameters are the parameters after adjustment.
After the data of each day of each time sequence are updated in the step (5), calculating whether a conversion inflection point appears on the current day of the time sequence by combining the data of the current day; meanwhile, a binary search method is used for predicting the condition of the occurrence of the inflection point of the switching in the tomorrow, and the method specifically comprises the following steps: 2 types of transformed inflection point prediction maps were generated each day: a conversion inflection point diagram in a high-frequency mode and a conversion inflection point diagram in a low-frequency mode. For a certain day of a general time sequence, an upper inflection point or a lower inflection point is not necessarily available, and most of the time is a stage of prediction judgment invariance;
but the condition of the occurrence of the inflection point of the tomorrow is estimated by a binary search method at the same time every day; if the current state is a continuous stage after the upper inflection point, estimating the condition of the lower inflection point appearing on the next day; and conversely, if the current state is a continuous stage after the lower inflection point, estimating the condition that the upper inflection point appears on the next day. Aiming at time sequences in different application fields, such as air temperature time sequences, e-commerce sales time sequences and the like, the change range of the sequence value is limited, so that when an upper inflection point or a lower inflection point possibly appears in the predicted tomorrow, the change range is obtained in a binary search method mode. Taking the predicted inflection point condition as an example, the following details are provided:
(1) adding a new data D at the end of the time series data0The initial value is the upper limit value of the current change amplitude; calculating whether the new data is an upper inflection point, if not, keeping the original state; if yes, entering the next step;
(2) the value interval of the variation amplitude is halved, and the new data D is calculated by the median of the interval0. Calculating the upper inflection point by the new sequence, calculating a new variation range according to whether the last bit is the upper inflection point mark, and calculating new data D0Until new data D is calculated0The values remain unchanged. At this time D0I.e. the condition for the occurrence of an up-comer.
Drawings
FIG. 1 is a flow chart of a time series transition inflection point calculation method of the present invention. Including the design of computational model methods and parameter training processes.
Fig. 2 illustrates a conversion inflection point diagram of 2 types generated by using a stock time series application as an example. The upper part is a conversion inflection point diagram under the high-frequency operation mode, and the lower part is a conversion inflection point diagram under the low-frequency operation mode. The application results of the temperature time sequence, the electroencephalogram time sequence and the like can be obtained.
Detailed Description
The invention is described in detail below with reference to the figures and examples.
The invention discloses a method for calculating a trend conversion inflection point of a time sequence. The method is a technical data analysis method, calculation and analysis are carried out after daily data updating, no future function exists, and therefore the generation of subsequent conversion inflection points cannot influence the result generated in the prior art. Assuming the time series S, the specific steps for performing the conversion inflection point calculation based on the current and historical data thereof are as follows.
Firstly, designing a conversion inflection point calculation framework based on a long-short term line.
The basic computing framework is: two trend curves are calculated based on current and historical data: long term and short term lines. When the short-term line passes through the long-term line from bottom to top, recording the intersection point as an upper inflection point; conversely, when the short-term line passes through the long-term line from top to bottom, the intersection point is recorded as a lower inflection point.
The long and short term lines are defined in a uniform manner as:
Line = EMA(SLOPE(D,K)*T+D,L)
the variables in the formula are explained below:
-D: representing the input time series data, which may be time series values of the day, sequence statistics, etc., or some combination of transformations between the series data, etc.;
-K: the method is characterized in that the method is a parameter variable and takes time series data of near K time periods as a calculation basis;
-SLOPE (D, K): representing time series data D in nearly K time periods, calculating the slope of a linear regression line;
-T: represents the forward T steps;
-SLOPE (D, K) × T + D: representing new time sequence data after the T steps are carried out forwards through the slope trend inertia of the regression line of the near K days;
- EMA (P, L): the exponential average index, i.e. the moving average with decreasing weighting in exponential form is performed on the sequence P to be weighted. The method specifically comprises the following steps: EMAtoday=α * Ptoday + ( 1 - α ) * EMAyesterday;
Wherein α is a smoothing index, and is taken as 2/(L + 1). When the formula is recurred continuously until EMA1EMA1Is undefined when EMA is present1Direct value of P1。
According to the difference of the values of the parameter variables K, T, L, the long and short lines or the buying and selling lines are further defined in a unified mode. As defined typically below.
Short-term line: ShortLine = EMA (SLOPE (D, K)b)*Tb+D,Lb) = EMA (SLOPE (D,8) × 4+ D,14), i.e. after 4 steps forward with a fitted SLOPE of approximately 8 days as the trend inertia, an exponentially weighted moving average with a parameter of 14 is performed. In the formula Kb =8, Tb =4, Lb =14。
Long-term line: LongLine = EMA (SLOPE (D, K)s)*Ts+D,Ls) = EMA (SLOPE (D,20) × 10+ D,35), i.e. after moving 10 steps forward with a fitted SLOPE of approximately 20 days as the trend inertia, an exponentially weighted moving average with a parameter of 35 is performed. In the formula Ks =20,Ts=10,Ls=35。
And secondly, adjusting parameters for generating the long-term line based on the slope combination.
Depth detail adjustment is performed on the long-term line based on the above framework and definition of the long-term and short-term lines. SLOPE (D, K) that primarily tunes long-term line LongLines)*TsAnd (4) partial.
The adjustment to the long-term line based on the slope combination is as follows:
the main objective is to adjust TsI.e. the number of steps taken forward. Note that the recent SLOPE SLP1= SLOPE (D,20), the short-term SLOPE SLP2= SLOPE (D,4), and ABS represents taking absolute values, and then there are four SLOPE combinations:
(1) SLP1>0 and SLP2>0. At this time, if SLP2>= SLP1, then reduce TsA value of (d);
(2) SLP1>0 and SLP2<0. At this time if ABS (SLP2)>= SLP1, increase TsA value of (d);
(3) SLP1<0 and SLP2>0. At this time, if SLP2>= ABS (SLP1), increase TsA value of (d);
(4) SLP1<0 and SLP2<0. At this time if ABS (SLP2)>ABS (SLP1), then T is decreasedsA value of (d);
other cases TsRemain unchanged.
And thirdly, dividing the output result into two typical modes of high frequency and low frequency according to the range of the calculation data of the short-term line.
After the long term line is determined, the short term line is processed next.
According to the foregoing definition, the short term line is: ShortLine = EMA (SLOPE (D, K)b)*Tb+D,Lb) Further defining two groups of phases according to the range of the calculated data of the short term lineThe relevant parameters are as follows: ultra-short term and short term.
Ultra-short term line: ShortLineA = EMA (SLOPE (D,4) × 2+ D,7), when K is presentb =4, Tb =2, LbAnd =7, namely, the data move forwards by 2 steps according to the slope inertia of the data of nearly 4 days.
Short-term line: ShortLineB = EMA (SLOPE (D,8) × 4+ D,14), when K is presentb =8, Tb =4, Lb=14, i.e. 4 steps forward by the slope inertia of the data of nearly 8 days.
Combining the ultra-short term line and the short term line with the long term line, calculating an intersection point to obtain a conversion inflection point, and combining to form two modes: a high frequency mode and a low frequency mode.
High-frequency mode: short-term line = ShortLineA; long-term line = LongLine;
low-frequency mode: short-term line = ShortLineB; long-term line = LongLine.
And fourthly, combining historical data, generating a real prediction record according to the conversion inflection point, and calculating and increasing to perform parameter tuning.
The above short-term line ShortLine = EMA (SLOPE (D, K)b)*Tb+D,Lb) And long-term line LongLine = EMA (SLOPE (D, K)s)*Ts+D,Ls) The definitions relate to parameters including Kb，Tb，Lb，Ks，Ts，LsAnd the like, in the foregoing description, values of some example parameters are given for convenience of description. In actual application, parameter optimization is performed by combining historical data of each time series, and the specific steps are as follows.
4.1 take a time series as an example, take data of a period of time. Typically only the data of the last year or half year is taken as training data. The acquired time-series data is input as a parameter D. The parameter D is a sequence data, or a statistic or combination of sequence data.
4.2 adjusting only one parameter, fixing other parameters and generating a conversion inflection point. Since the adjustment cannot be performed simultaneously due to the inclusion of a plurality of parameters, other parameters are fixed first, and only one parameter is left for adjustment. Generally, the value of this parameter is an integer value, and only a few of the countable values can be taken. After the parameters are valued, a short-term line and a long-term line are generated, the intersection point of the short-term line and the long-term line is calculated, and an upper inflection point and a lower inflection point are further determined according to whether the short-term line passes through the long-term line or passes through the long-term line.
4.3 generating a real prediction record according to the conversion inflection point, and recording the increase of the section from the occurrence of the upper inflection point to the occurrence of the lower inflection point.
4.4 calculate the number of transition corners and the total growth. And (4) counting the increasing condition of each conversion inflection point, and summing the increasing condition after converting the number of the inflection points into cost to obtain a final increasing result.
4.5 comparing with the last growth result, retaining the result and parameter of large growth.
And 4.6, repeatedly executing the steps 4.2-4.5 to obtain the optimal value of the current adjusting parameter.
4.7 fixing the value of the tuning parameter, and exchanging the value of another parameter in a similar procedure. After all parameters have been adjusted, this is counted as one round. And (4) carrying out several rounds of parameter adjustment according to the steps, wherein finally obtained parameters are the parameters after adjustment.
Fifthly, after data of each time sequence are updated every day, calculating whether a conversion inflection point appears on the current day of the time sequence by combining the data of the current day; meanwhile, the condition of the inflection point of the switching on the tomorrow is estimated by a binary search method.
The judgment of the time series transition inflection point is performed after the daily data update. Based on the tuning parameters obtained in the previous step, 2 types of conversion inflection point prediction graphs are generated every day: a conversion inflection point diagram in a high-frequency mode and a conversion inflection point diagram in a low-frequency mode. For a certain day of a general time sequence, there is not necessarily an upper inflection point or a lower inflection point, and most of the time is a stage of unchanging prediction judgment.
But the condition of the inflection point of the switching on the next day is estimated by a binary search method at the same time every day.
And estimating the condition of the inflection point of the switching on the tomorrow for the time sequence. If the current state is a continuous stage after the upper inflection point, estimating the condition of the lower inflection point appearing on the next day; and conversely, if the current state is a continuous stage after the lower inflection point, estimating the condition that the upper inflection point appears on the next day. Aiming at time sequences in different application fields, such as air temperature time sequences, e-commerce sales time sequences and the like, the change range of the sequence value is limited, so that when an upper inflection point or a lower inflection point possibly appears in the predicted tomorrow, the change range is obtained in a binary search method mode.
Taking the predicted inflection point condition as an example, the details are as follows.
(1) Adding a new data D at the end of the time series data0The initial value is the upper limit value of the current change amplitude; calculating whether the new data is an upper inflection point, if not, keeping the original state; if yes, the next step is carried out.
(2) The value interval of the variation amplitude is halved, and the new data D is calculated by the median of the interval0. Calculating the upper inflection point by the new sequence, calculating a new variation range according to whether the last bit is the upper inflection point mark, and calculating new data D0Until new data D is calculated0The values remain unchanged. At this time D0I.e. the condition for the occurrence of an up-comer.
In summary, the invention discloses a method for calculating a trend transformation inflection point of a time series. A frame for calculating the inflection point of time series conversion is designed, and related parameters can be adjusted according to historical trends and summary experiences of the time series under the frame. The method divides the output result into two typical modes: the system comprises a high-frequency mode and a low-frequency mode, wherein the two modes are trained respectively by historical data to obtain optimal parameters. After the data of the time sequence is updated every day, whether a conversion inflection point appears on the time sequence on the day is calculated and predicted by combining the data on the day, and meanwhile, the condition that an upper inflection point or a lower inflection point appears on the next day is rapidly predicted by a binary search method.
The method is applied to the following time sequence data, including air temperature time sequence, electric company sales volume time sequence, electroencephalogram time sequence, electric quantity time sequence, stock time sequence, futures time sequence and the like. Although specific embodiments of the invention have been disclosed for illustrative purposes and the accompanying drawings, which are included to provide a further understanding of the invention and are incorporated by reference, those skilled in the art will appreciate that: no alterations, changes, and modifications are possible without departing from the spirit and scope of the invention, as defined in the appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments and the accompanying drawings. The presently disclosed embodiments are to be considered in all respects as illustrative and not restrictive on the scope of the appended claims.

Claims (6)

1. A time series trend transformation inflection point prediction calculation method and a device thereof are characterized in that the method comprises the following steps:
(1) designing a conversion inflection point calculation framework based on the long and short term lines;
(2) adjusting parameters for generating a long-term line based on a slope combination mode;
(3) according to the range of the calculated data of the short-term line, dividing the output result into two typical modes of high frequency and low frequency;
(4) generating a real prediction record according to the conversion inflection point by combining historical data, and calculating increase to perform parameter tuning;
(5) after the data of the time sequence are updated every day, calculating whether a conversion inflection point appears on the current day of the time sequence by combining the data of the current day; meanwhile, the condition of the inflection point of the switching on the tomorrow is estimated by a binary search method.
2. The method and the device for predicting and calculating inflection points of trend transitions of time series according to claim 1, wherein in the computation framework of inflection points of transitions designed in step 1, the long and short term lines are defined by Line = EMA (SLOPE (D, K) × T + D, L) in a unified manner, and the unified idea is that new data is formed after the inertia goes forward by T steps, and the new data is subjected to exponential weighting of L parameters; then, based on a unified form, according to the difference of values of the parameter variable K, T, L, defining a short-term line and a long-term line:
short-term wire shortlink = EMA (SLOPE (D, K)b)*Tb+D,Lb) = EMA(SLOPE(D,8)*4+D,14)；
Long-term line LongLine= EMA(SLOPE(D,Ks)*Ts+D,Ls) = EMA(SLOPE(D,20)*10+D,35)。
3. The time-series trend transition inflection point prediction method and device as claimed in claim 1, wherein the adjustment of the long-term line parameters in step 2 is mainly aimed at adjusting TsBut there may be many ways including, but not limited to, the way in which the long-term line is adjusted based on a combination of recent and short-term slopes in the invention.
4. The method and apparatus as claimed in claim 1, wherein the step 3 further defines two sets of parameters, i.e. ultra-short term and short term, according to the range of the calculated data of the short term line: ultrashort line shortline a = EMA (SLOPE (D,4) × 2+ D,7), short-term line shortline b = EMA (SLOPE (D,8) × 4+ D, 14); and the two groups of parameters are combined with the LongLine to form two modes, namely a high-frequency mode: short-term line = ShortLineA; long-term line = LongLine; low-frequency mode: short-term line = ShortLineB; long-term line = LongLine.
5. The time series trend transition inflection point prediction method and device as claimed in claim 1, wherein the parameter tuning process of step 4 is a generalized gradient descent method; firstly, fixing other parameters and only adjusting one parameter; after the parameter is adjusted, adjusting other parameters by similar steps; after multiple rounds of adjustment, better parameters can be obtained; in the adjusting process, a real prediction record is generated by the conversion inflection point, the increase is calculated, and the total increase of the conversion cost of the conversion inflection point times is taken as an objective function.
6. The method and apparatus for predicting inflection points in trend of time series according to claim 1, wherein in step 5, in addition to predicting whether the inflection point appears in the time series of each day, the condition of the inflection point appearing in the next day is predicted by binary search; aiming at the limit value of the change amplitude of the time sequence without application field, binary approximation is carried out on the critical point of the change inflection point in a mode of assuming the change amplitude value of the tomorrow and calculating the change inflection point.
CN202010919043.5A 2020-09-04 2020-09-04 Time series trend transformation inflection point prediction calculation method and device Pending CN111949713A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010919043.5A CN111949713A (en) 2020-09-04 2020-09-04 Time series trend transformation inflection point prediction calculation method and device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010919043.5A CN111949713A (en) 2020-09-04 2020-09-04 Time series trend transformation inflection point prediction calculation method and device

Publications (1)

Publication Number Publication Date
CN111949713A true CN111949713A (en) 2020-11-17

Family

ID=73356179

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010919043.5A Pending CN111949713A (en) 2020-09-04 2020-09-04 Time series trend transformation inflection point prediction calculation method and device

Country Status (1)

Country Link
CN (1) CN111949713A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112580805A (en) * 2020-12-25 2021-03-30 三星（中国）半导体有限公司 Method and device for quantizing neural network model

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112580805A (en) * 2020-12-25 2021-03-30 三星（中国）半导体有限公司 Method and device for quantizing neural network model

Similar Documents

Publication Publication Date Title
Ding et al. Forecasting China's electricity consumption using a new grey prediction model
Haq et al. A new hybrid model for short-term electricity load forecasting
KR101998553B1 (en) Prediction Method of Short-Term Wind Speed and Wind Power and Power Supply Line Voltage Prediction Method Therefore
JP5516390B2 (en) Quality prediction apparatus, quality prediction method, program, and computer-readable recording medium
JP5434837B2 (en) Quality prediction apparatus, quality prediction method, program, and computer-readable recording medium
CN113544599A (en) Method for executing a process and optimizing a control signal used in the process
CN110443417A (en) Multiple-model integration load forecasting method based on wavelet transformation
Dash et al. MDHS–LPNN: a hybrid FOREX predictor model using a Legendre polynomial neural network with a modified differential harmony search technique
Freitas et al. Model combination in neural-based forecasting
Dudek et al. Medium-term electric energy demand forecasting using Nadaraya-Watson estimator
CN111949713A (en) Time series trend transformation inflection point prediction calculation method and device
Deng et al. Offline reinforcement learning for industrial process control: A case study from steel industry
Wang et al. Granular-based multilayer spatiotemporal network with control gates for energy prediction of steel industry
Lei et al. A novel time-delay neural grey model and its applications
Ratakonda et al. Seasonal trend analysis on multi-variate time series data
Bebarta et al. Polynomial based functional link artificial recurrent neural network adaptive system for predicting Indian stocks
Araujo et al. Hybrid intelligent design of morphological-rank-linear perceptrons for software development cost estimation
JP2020154364A (en) Transformer cost prediction device
CN115829123A (en) Natural gas demand prediction method and device based on grey model and neural network
JP7060130B1 (en) Operation support equipment, operation support methods and programs
Pełka et al. Prediction of monthly electric energy consumption using pattern-based fuzzy nearest neighbour regression
Mohd Lip et al. Comparative study of smoothing methods and box-jenkins model in forecasting unemployment rate in Malaysia
KR20220109935A (en) Ansemble device for predicting the power
Yanchenko et al. Stanza: a nonlinear state space model for probabilistic inference in non-stationary time series
Li et al. A multi-output integration residual network for predicting time series data with diverse scales

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination